[BLIP-2] Improve conversion script (huggingface#24854)

* Improve conversion script

* Add int8 code example

* Update tip

* Fix code

* Fix code snippet

* Add nucleus sampling

* More improvements

* Address comments

* Address comments

src/transformers/models/blip_2/configuration_blip_2.py

@@ -90,7 +90,7 @@ def __init__(
         image_size=224,
         patch_size=14,
         hidden_act="gelu",
-        layer_norm_eps=0.00001,
+        layer_norm_eps=1e-6,
         attention_dropout=0.0,
         initializer_range=1e-10,
         qkv_bias=True,

src/transformers/models/blip_2/convert_blip_2_original_to_pytorch.py

@@ -24,7 +24,8 @@
 import torch
 
 # pip3 install salesforce-lavis
-# I'm actually installing a slightly modified version: pip3 install git+https://github.com/nielsrogge/LAVIS.git@fix_lavis
+# I'm actually installing a slightly modified version: pip3 install -U git+https://github.com/nielsrogge/LAVIS.git@blip2_float32
+# to make sure we can compare both original and HF implementation in float32
 from lavis.models import load_model_and_preprocess
 from PIL import Image
 
@@ -37,6 +38,7 @@
     BlipImageProcessor,
     OPTConfig,
     T5Config,
+    set_seed,
 )
 from transformers.utils.constants import OPENAI_CLIP_MEAN, OPENAI_CLIP_STD
 
@@ -145,11 +147,16 @@ def convert_blip2_checkpoint(model_name, pytorch_dump_folder_path=None, push_to_
 
     name, type = model_name_to_original[model_name]
 
+    # note: this script is tested on 2 GPUs, as models are compared in float32,
+    # which requires quite some memory. Hence loading both on a
+    # separate device is the easiest to compare
+    hf_model_device = "cuda:0" if torch.cuda.is_available() else "cpu"
+    lavis_device = "cuda:1" if torch.cuda.is_available() else "cpu"
+
     # load original model
     print("Loading original model...")
-    device = "cuda" if torch.cuda.is_available() else "cpu"
     original_model, vis_processors, _ = load_model_and_preprocess(
-        name=name, model_type=type, is_eval=True, device=device
+        name=name, model_type=type, is_eval=True, device=lavis_device
     )
     original_model.eval()
     print("Done!")
@@ -185,61 +192,53 @@ def convert_blip2_checkpoint(model_name, pytorch_dump_folder_path=None, push_to_
     assert unexpected_keys == ["qformer.embeddings.position_ids"]
 
     image = load_demo_image()
-    original_pixel_values = vis_processors["eval"](image).unsqueeze(0).to(device)
-    input_ids = tokenizer(["\n"], return_tensors="pt").input_ids.to(device)
+    original_pixel_values = vis_processors["eval"](image).unsqueeze(0).to(lavis_device)
+    input_ids = tokenizer(["\n"], return_tensors="pt").input_ids.to(hf_model_device)
 
     # create processor
     image_processor = BlipImageProcessor(
         size={"height": image_size, "width": image_size}, image_mean=OPENAI_CLIP_MEAN, image_std=OPENAI_CLIP_STD
     )
     processor = Blip2Processor(image_processor=image_processor, tokenizer=tokenizer)
-    pixel_values = processor(images=image, return_tensors="pt").pixel_values.to(device)
+    pixel_values = processor(images=image, return_tensors="pt").pixel_values.to(hf_model_device)
 
     # make sure processor creates exact same pixel values
-    assert torch.allclose(pixel_values, original_pixel_values)
+    assert torch.allclose(pixel_values, original_pixel_values.to(pixel_values.device))
 
-    original_model.to(device)
-    hf_model.to(device)
+    original_model.to(lavis_device)
+    hf_model.to(hf_model_device)
     with torch.no_grad():
         if "opt" in model_name:
             original_logits = original_model({"image": original_pixel_values, "text_input": [""]}).logits
-            logits = hf_model(original_pixel_values, input_ids).logits
+            logits = hf_model(pixel_values, input_ids).logits
         else:
             original_logits = original_model(
                 {"image": original_pixel_values, "text_input": ["\n"], "text_output": ["\n"]}
             ).logits
             labels = input_ids.masked_fill(input_ids == tokenizer.pad_token_id, -100)
-            logits = hf_model(original_pixel_values, input_ids, labels=labels).logits
+            logits = hf_model(pixel_values, input_ids, labels=labels).logits
 
     assert original_logits.shape == logits.shape
     print("First values of original logits:", original_logits[0, :3, :3])
     print("First values of HF logits:", logits[0, :3, :3])
 
     # assert values
-    if model_name == "blip2-flan-t5-xl":
-        expected_slice_logits = torch.tensor(
-            [[-41.5850, -4.4440, -8.9922], [-47.4322, -5.9143, -1.7340]], device=device
-        )
-        assert torch.allclose(logits[0, :3, :3], expected_slice_logits, atol=1e-4)
-    elif model_name == "blip2-flan-t5-xl-coco":
-        expected_slice_logits = torch.tensor(
-            [[-57.0109, -9.8967, -12.6280], [-68.6578, -12.7191, -10.5065]], device=device
-        )
-    else:
-        # cast to same type
-        target_dtype = logits.dtype
-        assert torch.allclose(original_logits.to(target_dtype), logits, atol=1e-2)
+    assert torch.allclose(original_logits.to(logits.device), logits, atol=1e-4)
     print("Looks ok!")
 
     print("Generating a caption...")
-    prompt = ""
-    input_ids = tokenizer(prompt, return_tensors="pt").input_ids.to(device)
+    prompt = "Question: what object is in this image? Answer:"
+    input_ids = tokenizer(prompt, return_tensors="pt").input_ids.to(hf_model_device)
+
+    set_seed(42)
 
-    original_outputs = original_model.generate({"image": original_pixel_values})
+    original_outputs = original_model.generate(
+        {"image": original_pixel_values, "prompt": prompt}, use_nucleus_sampling=True
+    )
     outputs = hf_model.generate(
-        original_pixel_values,
+        pixel_values,
         input_ids,
-        do_sample=False,
+        do_sample=True,
         num_beams=5,
         max_length=30,
         min_length=1,
@@ -248,10 +247,9 @@ def convert_blip2_checkpoint(model_name, pytorch_dump_folder_path=None, push_to_
         length_penalty=1.0,
         temperature=1,
     )
-    print("Original generation:", original_outputs)
-    prompt_length = input_ids.shape[1]
-    output_text = processor.batch_decode(outputs[:, prompt_length:], skip_special_tokens=True)
+    output_text = processor.batch_decode(outputs, skip_special_tokens=True)
     output_text = [text.strip() for text in output_text]
+    print("Original generation:", original_outputs)
     print("HF generation:", output_text)
 
     if pytorch_dump_folder_path is not None:

src/transformers/models/blip_2/modeling_blip_2.py

@@ -1556,6 +1556,12 @@ def forward(
 
     One can optionally pass `input_ids` to the model, which serve as a text prompt, to make the language model continue
     the prompt. Otherwise, the language model starts generating text from the [BOS] (beginning-of-sequence) token.
+
+    <Tip>
+
+    Note that Flan-T5 checkpoints cannot be cast to float16. They are pre-trained using bfloat16.
+
+    </Tip>
     """,
     BLIP_2_START_DOCSTRING,
 )
@@ -1687,15 +1693,40 @@ def forward(
 
         >>> processor = Blip2Processor.from_pretrained("Salesforce/blip2-opt-2.7b")
         >>> model = Blip2ForConditionalGeneration.from_pretrained(
-        ...     "Salesforce/blip2-opt-2.7b", torch_dtype=torch.float16
-        ... )
-        >>> model.to(device)  # doctest: +IGNORE_RESULT
+        ...     "Salesforce/blip2-opt-2.7b", load_in_8bit=True, device_map={"": 0}, torch_dtype=torch.float16
+        ... )  # doctest: +IGNORE_RESULT
 
         >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
         >>> image = Image.open(requests.get(url, stream=True).raw)
 
         >>> prompt = "Question: how many cats are there? Answer:"
-        >>> inputs = processor(images=image, text=prompt, return_tensors="pt").to(device, torch.float16)
+        >>> inputs = processor(images=image, text=prompt, return_tensors="pt").to(device="cuda", dtype=torch.float16)
+
+        >>> generated_ids = model.generate(**inputs)
+        >>> generated_text = processor.batch_decode(generated_ids, skip_special_tokens=True)[0].strip()
+        >>> print(generated_text)
+        two
+        ```
+
+        Note that int8 inference is also supported through [bitsandbytes](https://github.com/TimDettmers/bitsandbytes).
+        This greatly reduces the amount of memory used by the model while maintaining the same performance.
+
+        ```python
+        >>> from PIL import Image
+        >>> import requests
+        >>> from transformers import Blip2Processor, Blip2ForConditionalGeneration
+        >>> import torch
+
+        >>> processor = Blip2Processor.from_pretrained("Salesforce/blip2-flan-t5-xl")
+        >>> model = Blip2ForConditionalGeneration.from_pretrained(
+        ...     "Salesforce/blip2-flan-t5-xl", load_in_8bit=True, device_map={"": 0}, torch_dtype=torch.bfloat16
+        ... )  # doctest: +IGNORE_RESULT
+
+        >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
+        >>> image = Image.open(requests.get(url, stream=True).raw)
+
+        >>> prompt = "Question: how many cats are there? Answer:"
+        >>> inputs = processor(images=image, text=prompt, return_tensors="pt").to(device="cuda", dtype=torch.bfloat16)
 
         >>> generated_ids = model.generate(**inputs)
         >>> generated_text = processor.batch_decode(generated_ids, skip_special_tokens=True)[0].strip()

utils/documentation_tests.txt

@@ -72,6 +72,7 @@ src/transformers/models/blip/image_processing_blip.py
 src/transformers/models/blip/modeling_blip.py
 src/transformers/models/blip/modeling_tf_blip.py
 src/transformers/models/blip/processing_blip.py
+src/transformers/models/blip_2/modeling_blip_2.py
 src/transformers/models/blip_2/processing_blip_2.py
 src/transformers/models/bloom/configuration_bloom.py
 src/transformers/models/bloom/tokenization_bloom_fast.py